lustre/osp/osp_sync.c

   1 /*
   2  * GPL HEADER START
   3  *
   4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   5  *
   6  * This program is free software; you can redistribute it and/or modify
   7  * it under the terms of the GNU General Public License version 2 only,
   8  * as published by the Free Software Foundation.
   9  *
  10  * This program is distributed in the hope that it will be useful, but
  11  * WITHOUT ANY WARRANTY; without even the implied warranty of
  12  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13  * General Public License version 2 for more details (a copy is included
  14  * in the LICENSE file that accompanied this code).
  15  *
  16  * You should have received a copy of the GNU General Public License
  17  * version 2 along with this program; If not, see
  18  * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
  19  *
  20  * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
  21  * CA 95054 USA or visit www.sun.com if you need additional information or
  22  * have any questions.
  23  *
  24  * GPL HEADER END
  25  */
  26 /*
  27  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  28  * Use is subject to license terms.
  29  *
  30  * Copyright (c) 2012, 2013, Intel Corporation.
  31  */
  32 /*
  33  * This file is part of Lustre, http://www.lustre.org/
  34  * Lustre is a trademark of Sun Microsystems, Inc.
  35  *
  36  * lustre/osp/osp_sync.c
  37  *
  38  * Lustre OST Proxy Device
  39  *
  40  * Author: Alex Zhuravlev <alexey.zhuravlev@intel.com>
  41  * Author: Mikhail Pershin <mike.pershin@intel.com>
  42  */
  43
  44 #define DEBUG_SUBSYSTEM S_MDS
  45
  46 #include <lustre_log.h>
  47 #include "osp_internal.h"
  48
  49 static int osp_sync_id_traction_init(struct osp_device *d);
  50 static void osp_sync_id_traction_fini(struct osp_device *d);
  51 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id);
  52 static void osp_sync_remove_from_tracker(struct osp_device *d);
  53
  54 /*
  55  * this is a components of OSP implementing synchronization between MDS and OST
  56  * it llogs all interesting changes (currently it's uig/gid change and object
  57  * destroy) atomically, then makes sure changes hit OST storage
  58  *
  59  * we have 4 queues of work:
  60  *
  61  * the first queue is llog itself, once read a change is stored in 2nd queue
  62  * in form of RPC (but RPC isn't fired yet).
  63  *
  64  * the second queue (opd_syn_waiting_for_commit) holds changes awaiting local
  65  * commit. once change is committed locally it migrates onto 3rd queue.
  66  *
  67  * the third queue (opd_syn_committed_here) holds changes committed locally,
  68  * but not sent to OST (as the pipe can be full). once pipe becomes non-full
  69  * we take a change from the queue and fire corresponded RPC.
  70  *
  71  * once RPC is reported committed by OST (using regular last_committed mech.)
  72  * the change jumps into 4th queue (opd_syn_committed_there), now we can
  73  * cancel corresponded llog record and release RPC
  74  *
  75  * opd_syn_changes is a number of unread llog records (to be processed).
  76  * notice this number doesn't include llog records from previous boots.
  77  * with OSP_SYN_THRESHOLD we try to batch processing a bit (TO BE IMPLEMENTED)
  78  *
  79  * opd_syn_rpc_in_progress is a number of requests in 2-4 queues.
  80  * we control this with OSP_MAX_IN_PROGRESS so that OSP don't consume
  81  * too much memory -- how to deal with 1000th OSTs ? batching could help?
  82  *
  83  * opd_syn_rpc_in_flight is a number of RPC in flight.
  84  * we control this with OSP_MAX_IN_FLIGHT
  85  */
  86
  87 /* XXX: do math to learn reasonable threshold
  88  * should it be ~ number of changes fitting bulk? */
  89
  90 #define OSP_SYN_THRESHOLD       10
  91 #define OSP_MAX_IN_FLIGHT       8
  92 #define OSP_MAX_IN_PROGRESS     4096
  93
  94 #define OSP_JOB_MAGIC           0x26112005
  95
  96 static inline int osp_sync_running(struct osp_device *d)
  97 {
  98         return !!(d->opd_syn_thread.t_flags & SVC_RUNNING);
  99 }
 100
 101 static inline int osp_sync_stopped(struct osp_device *d)
 102 {
 103         return !!(d->opd_syn_thread.t_flags & SVC_STOPPED);
 104 }
 105
 106 static inline int osp_sync_has_new_job(struct osp_device *d)
 107 {
 108         return ((d->opd_syn_last_processed_id < d->opd_syn_last_used_id) &&
 109                 (d->opd_syn_last_processed_id < d->opd_syn_last_committed_id))
 110                 || (d->opd_syn_prev_done == 0);
 111 }
 112
 113 static inline int osp_sync_low_in_progress(struct osp_device *d)
 114 {
 115         return d->opd_syn_rpc_in_progress < d->opd_syn_max_rpc_in_progress;
 116 }
 117
 118 static inline int osp_sync_low_in_flight(struct osp_device *d)
 119 {
 120         return d->opd_syn_rpc_in_flight < d->opd_syn_max_rpc_in_flight;
 121 }
 122
 123 static inline int osp_sync_has_work(struct osp_device *d)
 124 {
 125         /* has new/old changes and low in-progress? */
 126         if (osp_sync_has_new_job(d) && osp_sync_low_in_progress(d) &&
 127             osp_sync_low_in_flight(d) && d->opd_imp_connected)
 128                 return 1;
 129
 130         /* has remotely committed? */
 131         if (!list_empty(&d->opd_syn_committed_there))
 132                 return 1;
 133
 134         return 0;
 135 }
 136
 137 #define osp_sync_check_for_work(d)                      \
 138 {                                                       \
 139         if (osp_sync_has_work(d)) {                     \
 140                 wake_up(&d->opd_syn_waitq);    \
 141         }                                               \
 142 }
 143
 144 void __osp_sync_check_for_work(struct osp_device *d)
 145 {
 146         osp_sync_check_for_work(d);
 147 }
 148
 149 static inline int osp_sync_can_process_new(struct osp_device *d,
 150                                            struct llog_rec_hdr *rec)
 151 {
 152         LASSERT(d);
 153
 154         if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
 155                 return 0;
 156         if (!osp_sync_low_in_progress(d))
 157                 return 0;
 158         if (!osp_sync_low_in_flight(d))
 159                 return 0;
 160         if (!d->opd_imp_connected)
 161                 return 0;
 162         if (d->opd_syn_prev_done == 0)
 163                 return 1;
 164         if (d->opd_syn_changes == 0)
 165                 return 0;
 166         if (rec == NULL || rec->lrh_id <= d->opd_syn_last_committed_id)
 167                 return 1;
 168         return 0;
 169 }
 170
 171 int osp_sync_declare_add(const struct lu_env *env, struct osp_object *o,
 172                          llog_op_type type, struct thandle *th)
 173 {
 174         struct osp_thread_info  *osi = osp_env_info(env);
 175         struct osp_device       *d = lu2osp_dev(o->opo_obj.do_lu.lo_dev);
 176         struct llog_ctxt        *ctxt;
 177         int                      rc;
 178
 179         ENTRY;
 180
 181         /* it's a layering violation, to access internals of th,
 182          * but we can do this as a sanity check, for a while */
 183         LASSERT(th->th_dev == d->opd_storage);
 184
 185         switch (type) {
 186         case MDS_UNLINK64_REC:
 187                 osi->osi_hdr.lrh_len = sizeof(struct llog_unlink64_rec);
 188                 break;
 189         case MDS_SETATTR64_REC:
 190                 osi->osi_hdr.lrh_len = sizeof(struct llog_setattr64_rec);
 191                 break;
 192         default:
 193                 LBUG();
 194         }
 195
 196         /* we want ->dt_trans_start() to allocate per-thandle structure */
 197         th->th_tags |= LCT_OSP_THREAD;
 198
 199         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 200         LASSERT(ctxt);
 201
 202         rc = llog_declare_add(env, ctxt->loc_handle, &osi->osi_hdr, th);
 203         llog_ctxt_put(ctxt);
 204
 205         RETURN(rc);
 206 }
 207
 208 static int osp_sync_add_rec(const struct lu_env *env, struct osp_device *d,
 209                             const struct lu_fid *fid, llog_op_type type,
 210                             int count, struct thandle *th,
 211                             const struct lu_attr *attr)
 212 {
 213         struct osp_thread_info  *osi = osp_env_info(env);
 214         struct llog_ctxt        *ctxt;
 215         struct osp_txn_info     *txn;
 216         int                      rc;
 217
 218         ENTRY;
 219
 220         /* it's a layering violation, to access internals of th,
 221          * but we can do this as a sanity check, for a while */
 222         LASSERT(th->th_dev == d->opd_storage);
 223
 224         switch (type) {
 225         case MDS_UNLINK64_REC:
 226                 osi->osi_hdr.lrh_len = sizeof(osi->osi_unlink);
 227                 osi->osi_hdr.lrh_type = MDS_UNLINK64_REC;
 228                 osi->osi_unlink.lur_fid  = *fid;
 229                 osi->osi_unlink.lur_count = count;
 230                 break;
 231         case MDS_SETATTR64_REC:
 232                 rc = fid_to_ostid(fid, &osi->osi_oi);
 233                 LASSERT(rc == 0);
 234                 osi->osi_hdr.lrh_len = sizeof(osi->osi_setattr);
 235                 osi->osi_hdr.lrh_type = MDS_SETATTR64_REC;
 236                 osi->osi_setattr.lsr_oi  = osi->osi_oi;
 237                 LASSERT(attr);
 238                 osi->osi_setattr.lsr_uid = attr->la_uid;
 239                 osi->osi_setattr.lsr_gid = attr->la_gid;
 240                 osi->osi_setattr.lsr_valid = attr->la_valid;
 241                 break;
 242         default:
 243                 LBUG();
 244         }
 245
 246         txn = osp_txn_info(&th->th_ctx);
 247         LASSERT(txn);
 248
 249         txn->oti_current_id = osp_sync_id_get(d, txn->oti_current_id);
 250         osi->osi_hdr.lrh_id = txn->oti_current_id;
 251
 252         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
 253         if (ctxt == NULL)
 254                 RETURN(-ENOMEM);
 255         rc = llog_add(env, ctxt->loc_handle, &osi->osi_hdr, &osi->osi_cookie,
 256                       th);
 257         llog_ctxt_put(ctxt);
 258
 259         CDEBUG(D_OTHER, "%s: new record "DOSTID":%lu/%lu: %d\n",
 260                d->opd_obd->obd_name, POSTID(&osi->osi_cookie.lgc_lgl.lgl_oi),
 261                (unsigned long) osi->osi_cookie.lgc_lgl.lgl_ogen,
 262                (unsigned long) osi->osi_cookie.lgc_index, rc);
 263
 264         if (rc > 0)
 265                 rc = 0;
 266
 267         if (likely(rc == 0)) {
 268                 spin_lock(&d->opd_syn_lock);
 269                 d->opd_syn_changes++;
 270                 spin_unlock(&d->opd_syn_lock);
 271         }
 272
 273         RETURN(rc);
 274 }
 275
 276 int osp_sync_add(const struct lu_env *env, struct osp_object *o,
 277                  llog_op_type type, struct thandle *th,
 278                  const struct lu_attr *attr)
 279 {
 280         return osp_sync_add_rec(env, lu2osp_dev(o->opo_obj.do_lu.lo_dev),
 281                                 lu_object_fid(&o->opo_obj.do_lu), type, 1,
 282                                 th, attr);
 283 }
 284
 285 int osp_sync_gap(const struct lu_env *env, struct osp_device *d,
 286                  struct lu_fid *fid, int lost, struct thandle *th)
 287 {
 288         return osp_sync_add_rec(env, d, fid, MDS_UNLINK64_REC, lost, th, NULL);
 289 }
 290
 291 /*
 292  * it's quite obvious we can't maintain all the structures in the memory:
 293  * while OST is down, MDS can be processing thousands and thousands of unlinks
 294  * filling persistent llogs and in-core respresentation
 295  *
 296  * this doesn't scale at all. so we need basically the following:
 297  * a) destroy/setattr append llog records
 298  * b) once llog has grown to X records, we process first Y committed records
 299  *
 300  *  once record R is found via llog_process(), it becomes committed after any
 301  *  subsequent commit callback (at the most)
 302  */
 303
 304 /*
 305  * called for each atomic on-disk change (not once per transaction batch)
 306  * and goes over the list
 307  * XXX: should be optimized?
 308  */
 309
 310 /**
 311  * called for each RPC reported committed
 312  */
 313 static void osp_sync_request_commit_cb(struct ptlrpc_request *req)
 314 {
 315         struct osp_device *d = req->rq_cb_data;
 316
 317         CDEBUG(D_HA, "commit req %p, transno "LPU64"\n", req, req->rq_transno);
 318
 319         if (unlikely(req->rq_transno == 0))
 320                 return;
 321
 322         /* do not do any opd_dyn_rpc_* accounting here
 323          * it's done in osp_sync_interpret sooner or later */
 324
 325         LASSERT(d);
 326         LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 327         LASSERT(list_empty(&req->rq_exp_list));
 328
 329         ptlrpc_request_addref(req);
 330
 331         spin_lock(&d->opd_syn_lock);
 332         list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
 333         spin_unlock(&d->opd_syn_lock);
 334
 335         /* XXX: some batching wouldn't hurt */
 336         wake_up(&d->opd_syn_waitq);
 337 }
 338
 339 static int osp_sync_interpret(const struct lu_env *env,
 340                               struct ptlrpc_request *req, void *aa, int rc)
 341 {
 342         struct osp_device *d = req->rq_cb_data;
 343
 344         if (req->rq_svc_thread != (void *) OSP_JOB_MAGIC)
 345                 DEBUG_REQ(D_ERROR, req, "bad magic %p\n", req->rq_svc_thread);
 346         LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 347         LASSERT(d);
 348
 349         CDEBUG(D_HA, "reply req %p/%d, rc %d, transno %u\n", req,
 350                atomic_read(&req->rq_refcount),
 351                rc, (unsigned) req->rq_transno);
 352         LASSERT(rc || req->rq_transno);
 353
 354         if (rc == -ENOENT) {
 355                 /*
 356                  * we tried to destroy object or update attributes,
 357                  * but object doesn't exist anymore - cancell llog record
 358                  */
 359                 LASSERT(req->rq_transno == 0);
 360                 LASSERT(list_empty(&req->rq_exp_list));
 361
 362                 ptlrpc_request_addref(req);
 363
 364                 spin_lock(&d->opd_syn_lock);
 365                 list_add(&req->rq_exp_list, &d->opd_syn_committed_there);
 366                 spin_unlock(&d->opd_syn_lock);
 367
 368                 wake_up(&d->opd_syn_waitq);
 369         } else if (rc) {
 370                 struct obd_import *imp = req->rq_import;
 371                 /*
 372                  * error happened, we'll try to repeat on next boot ?
 373                  */
 374                 LASSERTF(req->rq_transno == 0 ||
 375                          req->rq_import_generation < imp->imp_generation,
 376                          "transno "LPU64", rc %d, gen: req %d, imp %d\n",
 377                          req->rq_transno, rc, req->rq_import_generation,
 378                          imp->imp_generation);
 379                 if (req->rq_transno == 0) {
 380                         /* this is the last time we see the request
 381                          * if transno is not zero, then commit cb
 382                          * will be called at some point */
 383                         LASSERT(d->opd_syn_rpc_in_progress > 0);
 384                         spin_lock(&d->opd_syn_lock);
 385                         d->opd_syn_rpc_in_progress--;
 386                         spin_unlock(&d->opd_syn_lock);
 387                 }
 388
 389                 wake_up(&d->opd_syn_waitq);
 390         } else if (d->opd_pre != NULL &&
 391                    unlikely(d->opd_pre_status == -ENOSPC)) {
 392                 /*
 393                  * if current status is -ENOSPC (lack of free space on OST)
 394                  * then we should poll OST immediately once object destroy
 395                  * is replied
 396                  */
 397                 osp_statfs_need_now(d);
 398         }
 399
 400         LASSERT(d->opd_syn_rpc_in_flight > 0);
 401         spin_lock(&d->opd_syn_lock);
 402         d->opd_syn_rpc_in_flight--;
 403         spin_unlock(&d->opd_syn_lock);
 404         if (unlikely(atomic_read(&d->opd_syn_barrier) > 0))
 405                 wake_up(&d->opd_syn_barrier_waitq);
 406         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 407                d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
 408                d->opd_syn_rpc_in_progress);
 409
 410         osp_sync_check_for_work(d);
 411
 412         return 0;
 413 }
 414
 415 /*
 416  * the function walks through list of committed locally changes
 417  * and send them to RPC until the pipe is full
 418  */
 419 static void osp_sync_send_new_rpc(struct osp_device *d,
 420                                   struct ptlrpc_request *req)
 421 {
 422         LASSERT(d->opd_syn_rpc_in_flight <= d->opd_syn_max_rpc_in_flight);
 423         LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 424
 425         ptlrpcd_add_req(req, PDL_POLICY_ROUND, -1);
 426 }
 427
 428 static struct ptlrpc_request *osp_sync_new_job(struct osp_device *d,
 429                                                struct llog_handle *llh,
 430                                                struct llog_rec_hdr *h,
 431                                                ost_cmd_t op,
 432                                                const struct req_format *format)
 433 {
 434         struct ptlrpc_request   *req;
 435         struct ost_body         *body;
 436         struct obd_import       *imp;
 437         int                      rc;
 438
 439         /* Prepare the request */
 440         imp = d->opd_obd->u.cli.cl_import;
 441         LASSERT(imp);
 442         req = ptlrpc_request_alloc(imp, format);
 443         if (req == NULL)
 444                 RETURN(ERR_PTR(-ENOMEM));
 445
 446         rc = ptlrpc_request_pack(req, LUSTRE_OST_VERSION, op);
 447         if (rc) {
 448                 ptlrpc_req_finished(req);
 449                 return ERR_PTR(rc);
 450         }
 451
 452         /*
 453          * this is a trick: to save on memory allocations we put cookie
 454          * into the request, but don't set corresponded flag in o_valid
 455          * so that OST doesn't interpret this cookie. once the request
 456          * is committed on OST we take cookie from the request and cancel
 457          */
 458         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 459         LASSERT(body);
 460         body->oa.o_lcookie.lgc_lgl = llh->lgh_id;
 461         body->oa.o_lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 462         body->oa.o_lcookie.lgc_index = h->lrh_index;
 463         INIT_LIST_HEAD(&req->rq_exp_list);
 464         req->rq_svc_thread = (void *) OSP_JOB_MAGIC;
 465
 466         req->rq_interpret_reply = osp_sync_interpret;
 467         req->rq_commit_cb = osp_sync_request_commit_cb;
 468         req->rq_cb_data = d;
 469
 470         ptlrpc_request_set_replen(req);
 471
 472         return req;
 473 }
 474
 475 static int osp_sync_new_setattr_job(struct osp_device *d,
 476                                     struct llog_handle *llh,
 477                                     struct llog_rec_hdr *h)
 478 {
 479         struct llog_setattr64_rec       *rec = (struct llog_setattr64_rec *)h;
 480         struct ptlrpc_request           *req;
 481         struct ost_body                 *body;
 482
 483         ENTRY;
 484         LASSERT(h->lrh_type == MDS_SETATTR64_REC);
 485
 486         /* lsr_valid can only be 0 or LA_UID/GID set */
 487         if (!rec->lsr_valid && !(rec->lsr_valid & ~(LA_UID | LA_GID))) {
 488                 CERROR("%s: invalid setattr record, lsr_valid:"LPU64"\n",
 489                        d->opd_obd->obd_name, rec->lsr_valid);
 490                 RETURN(-EINVAL);
 491         }
 492
 493         req = osp_sync_new_job(d, llh, h, OST_SETATTR, &RQF_OST_SETATTR);
 494         if (IS_ERR(req))
 495                 RETURN(PTR_ERR(req));
 496
 497         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 498         LASSERT(body);
 499         body->oa.o_oi = rec->lsr_oi;
 500         body->oa.o_uid = rec->lsr_uid;
 501         body->oa.o_gid = rec->lsr_gid;
 502         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 503         /* old setattr record (prior 2.6.0) doesn't have 'valid' stored,
 504          * we assume that both UID and GID are valid in that case. */
 505         if (rec->lsr_valid == 0) {
 506                 body->oa.o_valid |= (OBD_MD_FLUID | OBD_MD_FLGID);
 507         } else {
 508                 if (rec->lsr_valid & LA_UID)
 509                         body->oa.o_valid |= OBD_MD_FLUID;
 510                 if (rec->lsr_valid & LA_GID)
 511                         body->oa.o_valid |= OBD_MD_FLGID;
 512         }
 513
 514         osp_sync_send_new_rpc(d, req);
 515         RETURN(0);
 516 }
 517
 518 /* Old records may be in old format, so we handle that too */
 519 static int osp_sync_new_unlink_job(struct osp_device *d,
 520                                    struct llog_handle *llh,
 521                                    struct llog_rec_hdr *h)
 522 {
 523         struct llog_unlink_rec  *rec = (struct llog_unlink_rec *)h;
 524         struct ptlrpc_request   *req;
 525         struct ost_body         *body;
 526
 527         ENTRY;
 528         LASSERT(h->lrh_type == MDS_UNLINK_REC);
 529
 530         req = osp_sync_new_job(d, llh, h, OST_DESTROY, &RQF_OST_DESTROY);
 531         if (IS_ERR(req))
 532                 RETURN(PTR_ERR(req));
 533
 534         body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 535         LASSERT(body);
 536         ostid_set_seq(&body->oa.o_oi, rec->lur_oseq);
 537         ostid_set_id(&body->oa.o_oi, rec->lur_oid);
 538         body->oa.o_misc = rec->lur_count;
 539         body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID;
 540         if (rec->lur_count)
 541                 body->oa.o_valid |= OBD_MD_FLOBJCOUNT;
 542
 543         osp_sync_send_new_rpc(d, req);
 544         RETURN(0);
 545 }
 546
 547 static int osp_prep_unlink_update_req(const struct lu_env *env,
 548                                       struct osp_device *osp,
 549                                       struct llog_handle *llh,
 550                                       struct llog_rec_hdr *h,
 551                                       struct ptlrpc_request **reqp)
 552 {
 553         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 554         struct dt_update_request        *update = NULL;
 555         struct ptlrpc_request           *req;
 556         const char                      *buf;
 557         struct llog_cookie              lcookie;
 558         int                             size;
 559         int                             rc;
 560         ENTRY;
 561
 562         update = out_create_update_req(&osp->opd_dt_dev);
 563         if (IS_ERR(update))
 564                 RETURN(PTR_ERR(update));
 565
 566         /* This can only happens for unlink slave directory, so decrease
 567          * ref for ".." and "." */
 568         rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
 569                                NULL, NULL);
 570         if (rc != 0)
 571                 GOTO(out, rc);
 572
 573         rc = out_insert_update(env, update, OUT_REF_DEL, &rec->lur_fid, 0,
 574                                NULL, NULL);
 575         if (rc != 0)
 576                 GOTO(out, rc);
 577
 578         lcookie.lgc_lgl = llh->lgh_id;
 579         lcookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 580         lcookie.lgc_index = h->lrh_index;
 581         size = sizeof(lcookie);
 582         buf = (const char *)&lcookie;
 583
 584         rc = out_insert_update(env, update, OUT_DESTROY, &rec->lur_fid, 1,
 585                                &size, &buf);
 586         if (rc != 0)
 587                 GOTO(out, rc);
 588
 589         rc = out_prep_update_req(env, osp->opd_obd->u.cli.cl_import,
 590                                  update->dur_req, &req);
 591         if (rc != 0)
 592                 GOTO(out, rc);
 593
 594         INIT_LIST_HEAD(&req->rq_exp_list);
 595         req->rq_svc_thread = (void *)OSP_JOB_MAGIC;
 596
 597         req->rq_interpret_reply = osp_sync_interpret;
 598         req->rq_commit_cb = osp_sync_request_commit_cb;
 599         req->rq_cb_data = osp;
 600
 601         ptlrpc_request_set_replen(req);
 602         *reqp = req;
 603 out:
 604         if (update != NULL)
 605                 out_destroy_update_req(update);
 606
 607         RETURN(rc);
 608 }
 609
 610 static int osp_sync_new_unlink64_job(const struct lu_env *env,
 611                                      struct osp_device *d,
 612                                      struct llog_handle *llh,
 613                                      struct llog_rec_hdr *h)
 614 {
 615         struct llog_unlink64_rec        *rec = (struct llog_unlink64_rec *)h;
 616         struct ptlrpc_request           *req = NULL;
 617         struct ost_body                 *body;
 618         int                              rc;
 619
 620         ENTRY;
 621         LASSERT(h->lrh_type == MDS_UNLINK64_REC);
 622
 623         if (d->opd_connect_mdt) {
 624                 rc = osp_prep_unlink_update_req(env, d, llh, h, &req);
 625                 if (rc != 0)
 626                         RETURN(rc);
 627         } else {
 628                 req = osp_sync_new_job(d, llh, h, OST_DESTROY,
 629                                        &RQF_OST_DESTROY);
 630                 if (IS_ERR(req))
 631                         RETURN(PTR_ERR(req));
 632
 633                 body = req_capsule_client_get(&req->rq_pill, &RMF_OST_BODY);
 634                 if (body == NULL)
 635                         RETURN(-EFAULT);
 636                 rc = fid_to_ostid(&rec->lur_fid, &body->oa.o_oi);
 637                 if (rc < 0)
 638                         RETURN(rc);
 639                 body->oa.o_misc = rec->lur_count;
 640                 body->oa.o_valid = OBD_MD_FLGROUP | OBD_MD_FLID |
 641                                    OBD_MD_FLOBJCOUNT;
 642         }
 643         osp_sync_send_new_rpc(d, req);
 644         RETURN(0);
 645 }
 646
 647 static int osp_sync_process_record(const struct lu_env *env,
 648                                    struct osp_device *d,
 649                                    struct llog_handle *llh,
 650                                    struct llog_rec_hdr *rec)
 651 {
 652         struct llog_cookie       cookie;
 653         int                      rc = 0;
 654
 655         cookie.lgc_lgl = llh->lgh_id;
 656         cookie.lgc_subsys = LLOG_MDS_OST_ORIG_CTXT;
 657         cookie.lgc_index = rec->lrh_index;
 658
 659         if (unlikely(rec->lrh_type == LLOG_GEN_REC)) {
 660                 struct llog_gen_rec *gen = (struct llog_gen_rec *)rec;
 661
 662                 /* we're waiting for the record generated by this instance */
 663                 LASSERT(d->opd_syn_prev_done == 0);
 664                 if (!memcmp(&d->opd_syn_generation, &gen->lgr_gen,
 665                             sizeof(gen->lgr_gen))) {
 666                         CDEBUG(D_HA, "processed all old entries\n");
 667                         d->opd_syn_prev_done = 1;
 668                 }
 669
 670                 /* cancel any generation record */
 671                 rc = llog_cat_cancel_records(env, llh->u.phd.phd_cat_handle,
 672                                              1, &cookie);
 673
 674                 return rc;
 675         }
 676
 677         /*
 678          * now we prepare and fill requests to OST, put them on the queue
 679          * and fire after next commit callback
 680          */
 681
 682         /* notice we increment counters before sending RPC, to be consistent
 683          * in RPC interpret callback which may happen very quickly */
 684         spin_lock(&d->opd_syn_lock);
 685         d->opd_syn_rpc_in_flight++;
 686         d->opd_syn_rpc_in_progress++;
 687         spin_unlock(&d->opd_syn_lock);
 688
 689         switch (rec->lrh_type) {
 690         /* case MDS_UNLINK_REC is kept for compatibility */
 691         case MDS_UNLINK_REC:
 692                 rc = osp_sync_new_unlink_job(d, llh, rec);
 693                 break;
 694         case MDS_UNLINK64_REC:
 695                 rc = osp_sync_new_unlink64_job(env, d, llh, rec);
 696                 break;
 697         case MDS_SETATTR64_REC:
 698                 rc = osp_sync_new_setattr_job(d, llh, rec);
 699                 break;
 700         default:
 701                 CERROR("%s: unknown record type: %x\n", d->opd_obd->obd_name,
 702                        rec->lrh_type);
 703                 /* we should continue processing */
 704                 return 0;
 705         }
 706
 707         if (likely(rc == 0)) {
 708                 spin_lock(&d->opd_syn_lock);
 709                 if (d->opd_syn_prev_done) {
 710                         LASSERT(d->opd_syn_changes > 0);
 711                         LASSERT(rec->lrh_id <= d->opd_syn_last_committed_id);
 712                         /*
 713                          * NOTE: it's possible to meet same id if
 714                          * OST stores few stripes of same file
 715                          */
 716                         if (rec->lrh_id > d->opd_syn_last_processed_id) {
 717                                 d->opd_syn_last_processed_id = rec->lrh_id;
 718                                 wake_up(&d->opd_syn_barrier_waitq);
 719                         }
 720
 721                         d->opd_syn_changes--;
 722                 }
 723                 CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 724                        d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
 725                        d->opd_syn_rpc_in_progress);
 726                 spin_unlock(&d->opd_syn_lock);
 727         } else {
 728                 spin_lock(&d->opd_syn_lock);
 729                 d->opd_syn_rpc_in_flight--;
 730                 d->opd_syn_rpc_in_progress--;
 731                 spin_unlock(&d->opd_syn_lock);
 732         }
 733
 734         CDEBUG(D_HA, "found record %x, %d, idx %u, id %u: %d\n",
 735                rec->lrh_type, rec->lrh_len, rec->lrh_index, rec->lrh_id, rc);
 736         return rc;
 737 }
 738
 739 static void osp_sync_process_committed(const struct lu_env *env,
 740                                        struct osp_device *d)
 741 {
 742         struct obd_device       *obd = d->opd_obd;
 743         struct obd_import       *imp = obd->u.cli.cl_import;
 744         struct ost_body         *body;
 745         struct ptlrpc_request   *req, *tmp;
 746         struct llog_ctxt        *ctxt;
 747         struct llog_handle      *llh;
 748         struct list_head         list;
 749         int                      rc, done = 0;
 750
 751         ENTRY;
 752
 753         if (list_empty(&d->opd_syn_committed_there))
 754                 return;
 755
 756         /*
 757          * if current status is -ENOSPC (lack of free space on OST)
 758          * then we should poll OST immediately once object destroy
 759          * is committed.
 760          * notice: we do this upon commit as well because some backends
 761          * (like DMU) do not release space right away.
 762          */
 763         if (d->opd_pre != NULL && unlikely(d->opd_pre_status == -ENOSPC))
 764                 osp_statfs_need_now(d);
 765
 766         /*
 767          * now cancel them all
 768          * XXX: can we improve this using some batching?
 769          *      with batch RPC that'll happen automatically?
 770          * XXX: can we store ctxt in lod_device and save few cycles ?
 771          */
 772         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
 773         LASSERT(ctxt);
 774
 775         llh = ctxt->loc_handle;
 776         LASSERT(llh);
 777
 778         INIT_LIST_HEAD(&list);
 779         spin_lock(&d->opd_syn_lock);
 780         list_splice(&d->opd_syn_committed_there, &list);
 781         INIT_LIST_HEAD(&d->opd_syn_committed_there);
 782         spin_unlock(&d->opd_syn_lock);
 783
 784         list_for_each_entry_safe(req, tmp, &list, rq_exp_list) {
 785                 struct llog_cookie *lcookie = NULL;
 786
 787                 LASSERT(req->rq_svc_thread == (void *) OSP_JOB_MAGIC);
 788                 list_del_init(&req->rq_exp_list);
 789
 790                 if (d->opd_connect_mdt) {
 791                         struct object_update_request *ureq;
 792                         struct object_update *update;
 793                         ureq = req_capsule_client_get(&req->rq_pill,
 794                                                       &RMF_OUT_UPDATE);
 795                         LASSERT(ureq != NULL &&
 796                                 ureq->ourq_magic == UPDATE_REQUEST_MAGIC);
 797
 798                         /* 1st/2nd is for decref . and .., 3rd one is for
 799                          * destroy, where the log cookie is stored.
 800                          * See osp_prep_unlink_update_req */
 801                         update = object_update_request_get(ureq, 2, NULL);
 802                         LASSERT(update != NULL);
 803                         lcookie = object_update_param_get(update, 0, NULL);
 804                         LASSERT(lcookie != NULL);
 805                 } else {
 806                         body = req_capsule_client_get(&req->rq_pill,
 807                                                       &RMF_OST_BODY);
 808                         LASSERT(body);
 809                         lcookie = &body->oa.o_lcookie;
 810                 }
 811                 /* import can be closing, thus all commit cb's are
 812                  * called we can check committness directly */
 813                 if (req->rq_transno <= imp->imp_peer_committed_transno) {
 814                         rc = llog_cat_cancel_records(env, llh, 1, lcookie);
 815                         if (rc)
 816                                 CERROR("%s: can't cancel record: %d\n",
 817                                        obd->obd_name, rc);
 818                 } else {
 819                         DEBUG_REQ(D_HA, req, "not committed");
 820                 }
 821
 822                 ptlrpc_req_finished(req);
 823                 done++;
 824         }
 825
 826         llog_ctxt_put(ctxt);
 827
 828         LASSERT(d->opd_syn_rpc_in_progress >= done);
 829         spin_lock(&d->opd_syn_lock);
 830         d->opd_syn_rpc_in_progress -= done;
 831         spin_unlock(&d->opd_syn_lock);
 832         CDEBUG(D_OTHER, "%s: %d in flight, %d in progress\n",
 833                d->opd_obd->obd_name, d->opd_syn_rpc_in_flight,
 834                d->opd_syn_rpc_in_progress);
 835
 836         osp_sync_check_for_work(d);
 837
 838         /* wake up the thread if requested to stop:
 839          * it might be waiting for in-progress to complete */
 840         if (unlikely(osp_sync_running(d) == 0))
 841                 wake_up(&d->opd_syn_waitq);
 842
 843         EXIT;
 844 }
 845
 846 /*
 847  * this is where most of queues processing happens
 848  */
 849 static int osp_sync_process_queues(const struct lu_env *env,
 850                                    struct llog_handle *llh,
 851                                    struct llog_rec_hdr *rec,
 852                                    void *data)
 853 {
 854         struct osp_device       *d = data;
 855         int                      rc;
 856
 857         do {
 858                 struct l_wait_info lwi = { 0 };
 859
 860                 if (!osp_sync_running(d)) {
 861                         CDEBUG(D_HA, "stop llog processing\n");
 862                         return LLOG_PROC_BREAK;
 863                 }
 864
 865                 /* process requests committed by OST */
 866                 osp_sync_process_committed(env, d);
 867
 868                 /* if we there are changes to be processed and we have
 869                  * resources for this ... do now */
 870                 if (osp_sync_can_process_new(d, rec)) {
 871                         if (llh == NULL) {
 872                                 /* ask llog for another record */
 873                                 CDEBUG(D_HA, "%lu changes, %u in progress, %u in flight\n",
 874                                        d->opd_syn_changes,
 875                                        d->opd_syn_rpc_in_progress,
 876                                        d->opd_syn_rpc_in_flight);
 877                                 return 0;
 878                         }
 879
 880                         /*
 881                          * try to send, in case of disconnection, suspend
 882                          * processing till we can send this request
 883                          */
 884                         do {
 885                                 rc = osp_sync_process_record(env, d, llh, rec);
 886                                 /*
 887                                  * XXX: probably different handling is needed
 888                                  * for some bugs, like immediate exit or if
 889                                  * OSP gets inactive
 890                                  */
 891                                 if (rc) {
 892                                         CERROR("can't send: %d\n", rc);
 893                                         l_wait_event(d->opd_syn_waitq,
 894                                                      !osp_sync_running(d) ||
 895                                                      osp_sync_has_work(d),
 896                                                      &lwi);
 897                                 }
 898                         } while (rc != 0 && osp_sync_running(d));
 899
 900                         llh = NULL;
 901                         rec = NULL;
 902                 }
 903
 904                 if (d->opd_syn_last_processed_id == d->opd_syn_last_used_id)
 905                         osp_sync_remove_from_tracker(d);
 906
 907                 l_wait_event(d->opd_syn_waitq,
 908                              !osp_sync_running(d) ||
 909                              osp_sync_can_process_new(d, rec) ||
 910                              !list_empty(&d->opd_syn_committed_there),
 911                              &lwi);
 912         } while (1);
 913 }
 914
 915 /*
 916  * this thread runs llog_cat_process() scanner calling our callback
 917  * to process llog records. in the callback we implement tricky
 918  * state machine as we don't want to start scanning of the llog again
 919  * and again, also we don't want to process too many records and send
 920  * too many RPCs a time. so, depending on current load (num of changes
 921  * being synced to OST) the callback can suspend awaiting for some
 922  * new conditions, like syncs completed.
 923  *
 924  * in order to process llog records left by previous boots and to allow
 925  * llog_process_thread() to find something (otherwise it'd just exit
 926  * immediately) we add a special GENERATATION record on each boot.
 927  */
 928 static int osp_sync_thread(void *_arg)
 929 {
 930         struct osp_device       *d = _arg;
 931         struct ptlrpc_thread    *thread = &d->opd_syn_thread;
 932         struct l_wait_info       lwi = { 0 };
 933         struct llog_ctxt        *ctxt;
 934         struct obd_device       *obd = d->opd_obd;
 935         struct llog_handle      *llh;
 936         struct lu_env            env;
 937         int                      rc, count;
 938
 939         ENTRY;
 940
 941         rc = lu_env_init(&env, LCT_LOCAL);
 942         if (rc) {
 943                 CERROR("%s: can't initialize env: rc = %d\n",
 944                        obd->obd_name, rc);
 945                 RETURN(rc);
 946         }
 947
 948         spin_lock(&d->opd_syn_lock);
 949         thread->t_flags = SVC_RUNNING;
 950         spin_unlock(&d->opd_syn_lock);
 951         wake_up(&thread->t_ctl_waitq);
 952
 953         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
 954         if (ctxt == NULL) {
 955                 CERROR("can't get appropriate context\n");
 956                 GOTO(out, rc = -EINVAL);
 957         }
 958
 959         llh = ctxt->loc_handle;
 960         if (llh == NULL) {
 961                 CERROR("can't get llh\n");
 962                 llog_ctxt_put(ctxt);
 963                 GOTO(out, rc = -EINVAL);
 964         }
 965
 966         rc = llog_cat_process(&env, llh, osp_sync_process_queues, d, 0, 0);
 967         LASSERTF(rc == 0 || rc == LLOG_PROC_BREAK,
 968                  "%lu changes, %u in progress, %u in flight: %d\n",
 969                  d->opd_syn_changes, d->opd_syn_rpc_in_progress,
 970                  d->opd_syn_rpc_in_flight, rc);
 971
 972         /* we don't expect llog_process_thread() to exit till umount */
 973         LASSERTF(thread->t_flags != SVC_RUNNING,
 974                  "%lu changes, %u in progress, %u in flight\n",
 975                  d->opd_syn_changes, d->opd_syn_rpc_in_progress,
 976                  d->opd_syn_rpc_in_flight);
 977
 978         /* wait till all the requests are completed */
 979         count = 0;
 980         while (d->opd_syn_rpc_in_progress > 0) {
 981                 osp_sync_process_committed(&env, d);
 982
 983                 lwi = LWI_TIMEOUT(cfs_time_seconds(5), NULL, NULL);
 984                 rc = l_wait_event(d->opd_syn_waitq,
 985                                   d->opd_syn_rpc_in_progress == 0,
 986                                   &lwi);
 987                 if (rc == -ETIMEDOUT)
 988                         count++;
 989                 LASSERTF(count < 10, "%s: %d %d %sempty\n",
 990                          d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
 991                          d->opd_syn_rpc_in_flight,
 992                          list_empty(&d->opd_syn_committed_there) ? "" : "!");
 993
 994         }
 995
 996         llog_cat_close(&env, llh);
 997         rc = llog_cleanup(&env, ctxt);
 998         if (rc)
 999                 CERROR("can't cleanup llog: %d\n", rc);
1000 out:
1001         LASSERTF(d->opd_syn_rpc_in_progress == 0,
1002                  "%s: %d %d %sempty\n",
1003                  d->opd_obd->obd_name, d->opd_syn_rpc_in_progress,
1004                  d->opd_syn_rpc_in_flight,
1005                  list_empty(&d->opd_syn_committed_there) ? "" : "!");
1006
1007         thread->t_flags = SVC_STOPPED;
1008
1009         wake_up(&thread->t_ctl_waitq);
1010
1011         lu_env_fini(&env);
1012
1013         RETURN(0);
1014 }
1015
1016 static int osp_sync_llog_init(const struct lu_env *env, struct osp_device *d)
1017 {
1018         struct osp_thread_info  *osi = osp_env_info(env);
1019         struct lu_fid           *fid = &osi->osi_fid;
1020         struct llog_handle      *lgh = NULL;
1021         struct obd_device       *obd = d->opd_obd;
1022         struct llog_ctxt        *ctxt;
1023         int                     rc;
1024
1025         ENTRY;
1026
1027         LASSERT(obd);
1028
1029         /*
1030          * open llog corresponding to our OST
1031          */
1032         OBD_SET_CTXT_MAGIC(&obd->obd_lvfs_ctxt);
1033         obd->obd_lvfs_ctxt.dt = d->opd_storage;
1034
1035         if (d->opd_connect_mdt)
1036                 lu_local_obj_fid(fid, SLAVE_LLOG_CATALOGS_OID);
1037         else
1038                 lu_local_obj_fid(fid, LLOG_CATALOGS_OID);
1039
1040         rc = llog_osd_get_cat_list(env, d->opd_storage, d->opd_index, 1,
1041                                    &osi->osi_cid, fid);
1042         if (rc) {
1043                 CERROR("%s: can't get id from catalogs: rc = %d\n",
1044                        obd->obd_name, rc);
1045                 RETURN(rc);
1046         }
1047
1048         CDEBUG(D_INFO, "%s: Init llog for %d - catid "DOSTID":%x\n",
1049                obd->obd_name, d->opd_index,
1050                POSTID(&osi->osi_cid.lci_logid.lgl_oi),
1051                osi->osi_cid.lci_logid.lgl_ogen);
1052
1053         rc = llog_setup(env, obd, &obd->obd_olg, LLOG_MDS_OST_ORIG_CTXT, obd,
1054                         &osp_mds_ost_orig_logops);
1055         if (rc)
1056                 RETURN(rc);
1057
1058         ctxt = llog_get_context(obd, LLOG_MDS_OST_ORIG_CTXT);
1059         LASSERT(ctxt);
1060
1061         if (likely(logid_id(&osi->osi_cid.lci_logid) != 0)) {
1062                 rc = llog_open(env, ctxt, &lgh, &osi->osi_cid.lci_logid, NULL,
1063                                LLOG_OPEN_EXISTS);
1064                 /* re-create llog if it is missing */
1065                 if (rc == -ENOENT)
1066                         logid_set_id(&osi->osi_cid.lci_logid, 0);
1067                 else if (rc < 0)
1068                         GOTO(out_cleanup, rc);
1069         }
1070
1071         if (unlikely(logid_id(&osi->osi_cid.lci_logid) == 0)) {
1072                 rc = llog_open_create(env, ctxt, &lgh, NULL, NULL);
1073                 if (rc < 0)
1074                         GOTO(out_cleanup, rc);
1075                 osi->osi_cid.lci_logid = lgh->lgh_id;
1076         }
1077
1078         LASSERT(lgh != NULL);
1079         ctxt->loc_handle = lgh;
1080
1081         rc = llog_cat_init_and_process(env, lgh);
1082         if (rc)
1083                 GOTO(out_close, rc);
1084
1085         rc = llog_osd_put_cat_list(env, d->opd_storage, d->opd_index, 1,
1086                                    &osi->osi_cid, fid);
1087         if (rc)
1088                 GOTO(out_close, rc);
1089
1090         /*
1091          * put a mark in the llog till which we'll be processing
1092          * old records restless
1093          */
1094         d->opd_syn_generation.mnt_cnt = cfs_time_current();
1095         d->opd_syn_generation.conn_cnt = cfs_time_current();
1096
1097         osi->osi_hdr.lrh_type = LLOG_GEN_REC;
1098         osi->osi_hdr.lrh_len = sizeof(osi->osi_gen);
1099
1100         memcpy(&osi->osi_gen.lgr_gen, &d->opd_syn_generation,
1101                sizeof(osi->osi_gen.lgr_gen));
1102
1103         rc = llog_cat_add(env, lgh, &osi->osi_gen.lgr_hdr, &osi->osi_cookie);
1104         if (rc < 0)
1105                 GOTO(out_close, rc);
1106         llog_ctxt_put(ctxt);
1107         RETURN(0);
1108 out_close:
1109         llog_cat_close(env, lgh);
1110 out_cleanup:
1111         llog_cleanup(env, ctxt);
1112         RETURN(rc);
1113 }
1114
1115 static void osp_sync_llog_fini(const struct lu_env *env, struct osp_device *d)
1116 {
1117         struct llog_ctxt *ctxt;
1118
1119         ctxt = llog_get_context(d->opd_obd, LLOG_MDS_OST_ORIG_CTXT);
1120         if (ctxt != NULL)
1121                 llog_cat_close(env, ctxt->loc_handle);
1122         llog_cleanup(env, ctxt);
1123 }
1124
1125 /*
1126  * initializes sync component of OSP
1127  */
1128 int osp_sync_init(const struct lu_env *env, struct osp_device *d)
1129 {
1130         struct l_wait_info       lwi = { 0 };
1131         struct task_struct      *task;
1132         int                      rc;
1133
1134         ENTRY;
1135
1136         rc = osp_sync_id_traction_init(d);
1137         if (rc)
1138                 RETURN(rc);
1139
1140         /*
1141          * initialize llog storing changes
1142          */
1143         rc = osp_sync_llog_init(env, d);
1144         if (rc) {
1145                 CERROR("%s: can't initialize llog: rc = %d\n",
1146                        d->opd_obd->obd_name, rc);
1147                 GOTO(err_id, rc);
1148         }
1149
1150         /*
1151          * Start synchronization thread
1152          */
1153         d->opd_syn_max_rpc_in_flight = OSP_MAX_IN_FLIGHT;
1154         d->opd_syn_max_rpc_in_progress = OSP_MAX_IN_PROGRESS;
1155         spin_lock_init(&d->opd_syn_lock);
1156         init_waitqueue_head(&d->opd_syn_waitq);
1157         init_waitqueue_head(&d->opd_syn_barrier_waitq);
1158         init_waitqueue_head(&d->opd_syn_thread.t_ctl_waitq);
1159         INIT_LIST_HEAD(&d->opd_syn_committed_there);
1160
1161         task = kthread_run(osp_sync_thread, d, "osp-syn-%u-%u",
1162                            d->opd_index, d->opd_group);
1163         if (IS_ERR(task)) {
1164                 rc = PTR_ERR(task);
1165                 CERROR("%s: cannot start sync thread: rc = %d\n",
1166                        d->opd_obd->obd_name, rc);
1167                 GOTO(err_llog, rc);
1168         }
1169
1170         l_wait_event(d->opd_syn_thread.t_ctl_waitq,
1171                      osp_sync_running(d) || osp_sync_stopped(d), &lwi);
1172
1173         RETURN(0);
1174 err_llog:
1175         osp_sync_llog_fini(env, d);
1176 err_id:
1177         osp_sync_id_traction_fini(d);
1178         return rc;
1179 }
1180
1181 int osp_sync_fini(struct osp_device *d)
1182 {
1183         struct ptlrpc_thread *thread = &d->opd_syn_thread;
1184
1185         ENTRY;
1186
1187         thread->t_flags = SVC_STOPPING;
1188         wake_up(&d->opd_syn_waitq);
1189         wait_event(thread->t_ctl_waitq, thread->t_flags & SVC_STOPPED);
1190
1191         /*
1192          * unregister transaction callbacks only when sync thread
1193          * has finished operations with llog
1194          */
1195         osp_sync_id_traction_fini(d);
1196
1197         RETURN(0);
1198 }
1199
1200 static DEFINE_MUTEX(osp_id_tracker_sem);
1201 static struct list_head osp_id_tracker_list =
1202                 LIST_HEAD_INIT(osp_id_tracker_list);
1203
1204 static void osp_sync_tracker_commit_cb(struct thandle *th, void *cookie)
1205 {
1206         struct osp_id_tracker   *tr = cookie;
1207         struct osp_device       *d;
1208         struct osp_txn_info     *txn;
1209
1210         LASSERT(tr);
1211
1212         txn = osp_txn_info(&th->th_ctx);
1213         if (txn == NULL || txn->oti_current_id < tr->otr_committed_id)
1214                 return;
1215
1216         spin_lock(&tr->otr_lock);
1217         if (likely(txn->oti_current_id > tr->otr_committed_id)) {
1218                 CDEBUG(D_OTHER, "committed: %u -> %u\n",
1219                        tr->otr_committed_id, txn->oti_current_id);
1220                 tr->otr_committed_id = txn->oti_current_id;
1221
1222                 list_for_each_entry(d, &tr->otr_wakeup_list,
1223                                     opd_syn_ontrack) {
1224                         d->opd_syn_last_committed_id = tr->otr_committed_id;
1225                         wake_up(&d->opd_syn_waitq);
1226                 }
1227         }
1228         spin_unlock(&tr->otr_lock);
1229 }
1230
1231 static int osp_sync_id_traction_init(struct osp_device *d)
1232 {
1233         struct osp_id_tracker   *tr, *found = NULL;
1234         int                      rc = 0;
1235
1236         LASSERT(d);
1237         LASSERT(d->opd_storage);
1238         LASSERT(d->opd_syn_tracker == NULL);
1239         INIT_LIST_HEAD(&d->opd_syn_ontrack);
1240
1241         mutex_lock(&osp_id_tracker_sem);
1242         list_for_each_entry(tr, &osp_id_tracker_list, otr_list) {
1243                 if (tr->otr_dev == d->opd_storage) {
1244                         LASSERT(atomic_read(&tr->otr_refcount));
1245                         atomic_inc(&tr->otr_refcount);
1246                         d->opd_syn_tracker = tr;
1247                         found = tr;
1248                         break;
1249                 }
1250         }
1251
1252         if (found == NULL) {
1253                 rc = -ENOMEM;
1254                 OBD_ALLOC_PTR(tr);
1255                 if (tr) {
1256                         d->opd_syn_tracker = tr;
1257                         spin_lock_init(&tr->otr_lock);
1258                         tr->otr_dev = d->opd_storage;
1259                         tr->otr_next_id = 1;
1260                         tr->otr_committed_id = 0;
1261                         atomic_set(&tr->otr_refcount, 1);
1262                         INIT_LIST_HEAD(&tr->otr_wakeup_list);
1263                         list_add(&tr->otr_list, &osp_id_tracker_list);
1264                         tr->otr_tx_cb.dtc_txn_commit =
1265                                                 osp_sync_tracker_commit_cb;
1266                         tr->otr_tx_cb.dtc_cookie = tr;
1267                         tr->otr_tx_cb.dtc_tag = LCT_MD_THREAD;
1268                         dt_txn_callback_add(d->opd_storage, &tr->otr_tx_cb);
1269                         rc = 0;
1270                 }
1271         }
1272         mutex_unlock(&osp_id_tracker_sem);
1273
1274         return rc;
1275 }
1276
1277 static void osp_sync_id_traction_fini(struct osp_device *d)
1278 {
1279         struct osp_id_tracker *tr;
1280
1281         ENTRY;
1282
1283         LASSERT(d);
1284         tr = d->opd_syn_tracker;
1285         if (tr == NULL) {
1286                 EXIT;
1287                 return;
1288         }
1289
1290         osp_sync_remove_from_tracker(d);
1291
1292         mutex_lock(&osp_id_tracker_sem);
1293         if (atomic_dec_and_test(&tr->otr_refcount)) {
1294                 dt_txn_callback_del(d->opd_storage, &tr->otr_tx_cb);
1295                 LASSERT(list_empty(&tr->otr_wakeup_list));
1296                 list_del(&tr->otr_list);
1297                 OBD_FREE_PTR(tr);
1298                 d->opd_syn_tracker = NULL;
1299         }
1300         mutex_unlock(&osp_id_tracker_sem);
1301
1302         EXIT;
1303 }
1304
1305 /*
1306  * generates id for the tracker
1307  */
1308 static __u32 osp_sync_id_get(struct osp_device *d, __u32 id)
1309 {
1310         struct osp_id_tracker *tr;
1311
1312         tr = d->opd_syn_tracker;
1313         LASSERT(tr);
1314
1315         /* XXX: we can improve this introducing per-cpu preallocated ids? */
1316         spin_lock(&tr->otr_lock);
1317         if (unlikely(tr->otr_next_id <= d->opd_syn_last_used_id)) {
1318                 spin_unlock(&tr->otr_lock);
1319                 CERROR("%s: next %u, last synced %lu\n",
1320                        d->opd_obd->obd_name, tr->otr_next_id,
1321                        d->opd_syn_last_used_id);
1322                 LBUG();
1323         }
1324
1325         if (id == 0)
1326                 id = tr->otr_next_id++;
1327         if (id > d->opd_syn_last_used_id)
1328                 d->opd_syn_last_used_id = id;
1329         if (list_empty(&d->opd_syn_ontrack))
1330                 list_add(&d->opd_syn_ontrack, &tr->otr_wakeup_list);
1331         spin_unlock(&tr->otr_lock);
1332         CDEBUG(D_OTHER, "new id %u\n", (unsigned) id);
1333
1334         return id;
1335 }
1336
1337 static void osp_sync_remove_from_tracker(struct osp_device *d)
1338 {
1339         struct osp_id_tracker *tr;
1340
1341         tr = d->opd_syn_tracker;
1342         LASSERT(tr);
1343
1344         if (list_empty(&d->opd_syn_ontrack))
1345                 return;
1346
1347         spin_lock(&tr->otr_lock);
1348         list_del_init(&d->opd_syn_ontrack);
1349         spin_unlock(&tr->otr_lock);
1350 }
1351